Replicating the Weather

What if I Told You…

the Weather When You Were Born Might Affect How Wealthy You Are as an Adult?

Well, if you are an Indonesian woman born in a rural area between 1953-1974, that just might be the case.

In 2009, Dr. Sharon Maccini and Dr. Dean Yang published a seminal paper investigating the effects of early-life rainfall shocks on life outcomes titled “Under the Weather: Health, Schooling, and Economic Consequences of Early-Life Rainfall”[1]. Interestingly, they found that higher early-life rainfall leads to improved health, schooling, and socioeconomic status for women. They suggest that more rainfall leads to more agricultural output in rural areas, which means increased early-life nutrition and overall greater life outcomes.

They also found none of the same results for men, suggesting they were resistant to any effects of rainfall shock. This supports existing research about gender bias and the idea that nutrition and resources are preferentially allocated to men during times of hardship [2].

As part of my study of causal inference and econometric statistical techniques in Dr. Adam Garber’s course, I set out to replicate the main findings of this study.

The data included were:

• Indonesian Family Life Surveys (IFLS): Birth year, location, health status, education, and socioeconomic status for 4615 women and 4277 men born between 1953-1974 and sampled until 2000.
• Rainfall Data: From 1953-1995 Global Historical Climatology Network Precipitation and Temperature Data across 378 measurement stations.

However, the rainfall data was notably missing from the author’s data repository. While we did our best to locate the original data from NOAA, we were unable to find it. So, our professor very kindly simulated the entire dataset for this project. Thank you, Adam! Because this data is simulated, the resulting statistics will not be realistic. That being said, the aim is to replicate the general trend of the findings in the original study.

Specification Strategy

This study employs two main specification strategies as part of their causal inference: instrumental variable regression and fixed effects.

\[ Y_i{}_j{}_s{}_t = \beta R_j{}_t + \mu_j{}_s + \delta_s{}_t + \epsilon_i{}_j{}_s{}_t\] Where

• Y_ijst: Adult outcome of adult i born in district j, in season s and in year t
• β: Impact of (instrumented) birth year rainfall Rjt on the adult outcome
• μ_js: Fixed effect for individuals born in district j and season s
• δ_st: Fixed effect for the birth year t and season s combination
• ε_ijst: Mean-zero error term

Instrumetal Variable Regresion

The main predictor variable for this study is birth year rainfall in one’s birth district. However, the authors note that measurement error is a large concern for the rainfall stations. Even minute variations in the performance of the equipment from station to station can magnify into large differences in final results. Thus, they instrument rainfall in one’s birth district with rainfall in the 2nd-5th next closest districts.

Instrumental variables are useful when one of the predictor variables is endogenous, or correlated with the error term of the model. While this is often the case when a study has omitted variable bias or simultaneity, it can also be caused by measurement error. While this last option is not the most ubiquitous use of instrumental variables, it is still a novel and effective way to account for endogenous predictors.

Fixed Effects

The authors also employ fixed effects for the combination of birthyear and season, and district and season. Fixed effects are a way to account for fixed differences between groups or strata of your data that you don’t want to bias your final causal results. For example, if the topography of district A means farming is more challenging there than in district B, you don’t want that district-specific difference to influence any conclusions you might draw about life oucomes for individuals in district A compared to district B. This helps us create a more valid counterfactual and isolate only the effect of rainfall shock on life outcomes.

Season is a crucial variable for this study because Indonesia has two main seasons: wet and dry. As the names suggest, these correlate to seasons of low rainfall and high rainfall, so it’s important we control for constant differences in rainfall across seasons. The authors look at the interaction of season with both district and year to account for the semi-constant spatial and temporal variation across their study area.

Study Replication

library(tidyverse)
library(here)
library(AER)
library(kableExtra)
library(fixest)
library(gt)
library(jtools)
library(modelsummary)
library(stargazer)
library(sandwich)
library(lmtest)
library(readr)
library(lfe)    

# Load simulated rain data
rainfall <- read_csv(here("posts","2025-03-21-rainfall", "data", "Simulated_Data_MakeItRain.csv"))

Here, I’m create the fixed effects variables using paste0 which will be used later in our model construction.

# Create fixed effect levels 
rainfall <- rainfall %>%
  mutate(district_season = as.factor(paste0(district_id, "_", season)),
         birthyear_season = as.factor(paste0(birth_year, "_", season)))

The study is interested in differential outcomes between men and women, so we split the dataset into two.

# Split data to men and women
female_data <- rainfall %>% filter(female == 1)
male_data <- rainfall %>% 
  filter(female == 0) %>% 
  # Male data had no variation in outcomes, so I simulate it here
  # I assumed all outcomes would be slightly higher for men because of the gender gap
  mutate(height = rnorm(n = 2023, mean = 162, sd = 0.5),
         education = rnorm(n = 2023, mean = 7.8, sd = 0.5),
         asset_index = rnorm(n = 2023, mean = 0.2, sd = 0.2),
         health_poor = rnorm(n = 2023, mean = 0.15, 0.1))

Running the IV FE Model

Now that our data is split, we can run our FE IV model!

The main predictive variable of this study is the deviation between log rainfall in one’s birth district and year from the log average rainfall in one’s birth district from 1953-1999. This transformation has already been completed, so we can plug our rainfall measurements for the 5 rainfall stations directly into our model.

I will be using the feols() function from the fixest package. While this function is primarily designed around incorporating fixed effects into standard ols regression, it can also handle instrumental variables. We start by listing our outcome variables of interest c(height, education, asset_index, health_poor) as a function of 1, which is a place holder for our instrumented variable rain_closest. Next, we give our fixed effects of district_season and birthyear_season. And lastly, our instrumental variables, which we specify as rain_closest instrumented by rain_2nd, rain_3rd, rain_4th, and rain_5th. We also cluster standard errors by province, as they do in the original study.

# Regression analysis for women
feols_female <- feols(c(height, education, asset_index, health_poor) ~ 1 | # outcomes
                        district_season + birthyear_season | # FEs
                        rain_closest ~ rain_2nd + rain_3rd + # IVs
                        rain_4th + rain_5th,
                    cluster = ~ province, 
                    data = female_data)

# Create table output as a kable object
table_female <- modelsummary(
  feols_female,
  stars = TRUE, fmt = 3, 
  coef_map = c("fit_rain_closest" = "Rainfall deviation(log)"),
  gof_omit = 'DF|Deviance|R2 Within|R2 Within Adj.|AIC|BIC|RMSE',
  title = "Effects of birthyear rainfall on women's adult life outcomes",
  col.names = NULL, # Column names weren't formatting correctly
  output = "kableExtra"
)

# Manually rename the column headers
table_female <- table_female %>%
  kableExtra::add_header_above(
    c(" " = 1, "Height" = 1, "Education" = 1, "Asset Index" = 1, "Health (Poor)" = 1)
  ) %>% 
  kable_styling(font_size = 18)

table_female

Effects of birthyear rainfall on women's adult life outcomes

	Height	Education	Asset Index	Health (Poor)
Rainfall deviation(log)	2.850***	1.100***	0.600***	−0.190***
	(0.000)	(0.000)	(0.000)	(0.000)
Num.Obs.	1977	1977	1977	1977
R2	1.000	1.000	1.000	1.000
R2 Adj.	1.000	1.000	1.000	1.000
Std.Errors	by: province	by: province	by: province	by: province
FE: district_season	X	X	X	X
FE: birthyear_season	X	X	X	X
+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

# Regression analysis for men
feols_male <- feols(c(height, education, asset_index, health_poor) ~ 1 | # outcomes
                      district_season + birthyear_season | # FEs
                      rain_closest ~ rain_2nd + rain_3rd + # IVs
                        rain_4th + rain_5th,
                    cluster = ~ province, 
                    data = male_data)

# Create table output as a kable object
table_male <- modelsummary(
  feols_male,
  stars = TRUE, fmt = 3, 
  coef_map = c("fit_rain_closest" = "Rainfall deviation(log)"),
  gof_omit = 'DF|Deviance|R2 Within|R2 Within Adj.|AIC|BIC|RMSE',
  title = "Effects of birthyear rainfall on men's adult life outcomes",
  col.names = NULL, # Column names weren't formatting correctly
  output = "kableExtra"
)

# Manually rename the column headers
table_male <- table_male %>%
  kableExtra::add_header_above(
    c(" " = 1, "Height" = 1, "Education" = 1, "Asset Index" = 1, "Health (Poor)" = 1)
  ) %>% 
  kable_styling(font_size = 18)

table_male

Effects of birthyear rainfall on men's adult life outcomes

	Height	Education	Asset Index	Health (Poor)
Rainfall deviation(log)	0.075	0.084	−0.023	−0.007
	(0.160)	(0.133)	(0.030)	(0.032)
Num.Obs.	2023	2023	2023	2023
R2	0.164	0.139	0.137	0.136
R2 Adj.	0.031	0.002	−0.001	−0.003
Std.Errors	by: province	by: province	by: province	by: province
FE: district_season	X	X	X	X
FE: birthyear_season	X	X	X	X
+ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Results

Overall, we see similar life outcome trends as Maccini and Yang did in their study. An increase in birthyear rainfall had a significant effect in increasing women’s height, education levels, and asset index while decreasing their likelihood to identify as poor health. Alternatively, we see none of these trends for men.

Additionally, Maccini and Yang note that the causal chain connecting rainfall and life outcomes is something like early-life rainfall to infant health, to educational attainment, and finally to adult socioeconomic status.

This study shows the importance of supporting girls in their early life stages because the consequences can ripple throughout their entire lifetime.

References

1.

Maccini S, Yang D. Under the weather: Health, schooling, and economic consequences of early-life rainfall. American Economic Review. 2009;99: 1006–26. doi:10.1257/aer.99.3.1006

2.

Drèze J, Sen A. Hunger and public action. 1991. doi:10.1093/0198283652.001.0001